hairtuq

Joined 9 years ago

Karma

Submissions

Member since

Mar 29

View on Hacker News

Recent submissions

Comment · 1 year ago

The mapping after we have the leading 1's count can be done in 3 instructions (in 32-bit math) on either x86 or ARM:

    t = 0x2020c6 >> x
    return (t << (t & 31)) >> 27

Comment · 1 year ago

If you want to optimize it a little more, you can combine isLess(s, 0x20) and isChar('"') into isLess(s ^ (kOnes * 0x02), 0x21) (this works since '"' is 0x22).

Comment · 1 year ago

Thanks, fixed.

Comment · 1 year ago

The article leaves out optimizing is_leap_year. Here is an interesting version that is correct (in the Proleptic Gregorian calendar) for 0 <= year <= 102499 and takes only 3 instructions:

    ((year * 1073750999) & 3221352463) <= 126976

How this works is surprisingly complex, I am thinking of writing a whole blog post about it...

Comment · 1 year ago

Here is a cute variant that doesn't need lzcnt nor tables, but only works up to 99,999:

    (((x + 393206) & (x + 524188)) ^ ((x + 916504) & (x + 514288))) >> 17

This is for integer log 10, but could be adapted for number of digits. It needs a wrapper for 64 bit to invoke it multiple times, but most numbers in a JSON are small, so it might even be competitive; it needs only 4 cycles with enough instruction level parallelism.

I gathered this idea from the output of a superoptimizer, it was fun to figure out how it works. For spoilers, see [1].

[1] https://github.com/rust-lang/rust/blob/master/library/core/s...

Comment · 1 year ago

The author wonders:

> In theory at least, the compiler can see that rule only has 256 values and create a reduced version of ca1d_rule_apply for each value. Whether it actually does is not of much practical concern when the rendering code is the bottle neck. However it’s interesting to see if the compiler can deduce the best solution or whether anything trips it up.

The compiler is unlikely to get the optimal result here. The core of this is finding the best instruction sequence for a ternary boolean operation encoded in 8 bits; it's the same job needed for emulating the AVX512F "vpternlog" instruction. This can always be done in at most 5 instructions (or 4 if you have andnot/ornot/xornot), but it's not straightforward to do this. Here is some code that calculates optimal instruction sequences (by letting z3 do the heavy lifting): https://github.com/falk-hueffner/ternary-logic-optimization

Comment · 2 years ago

Comment · 3 years ago

I wrote a solver for Atomix (https://github.com/falk-hueffner/atomixer), a similar game. Technically, the only difference is that in Atomix the solution location is not predefined, although probably based on level design this game actually feels quite different. Since the solver just tries all solution location anyway, it would be quite simple to adapt it to this game.

Comment · 3 years ago

    mask = -1 << i;
    return (x & ~mask) | ((x >> 1) & mask);

but with xor we can do it in one less instruction:

    mask = -1 << i;
    return ((x ^ (x >> 1)) & mask) ^ x;

Comment · 4 years ago

It seems you're shifting the goal posts now. The fact that consensus criticisms are not as harsh as you'd like doesn't mean that Israel is "totally off limits". Discussions and criticisms are happening; and the fact that public opinion doesn't match yours can not only be explained by that opinion being suppressed, but also by it being wrong.

Comment · 4 years ago

I don't see that. For example, it has been the official position of the German government for at least 20 years to strongly criticize the Israeli settlements. There's regular critical reports in the press, endless TV discussions and so on. Can you name someone who has actually had to suffer negative consequences for reasonable criticism of Israel?

Comment · 4 years ago

I wrote an ingredient parser that is basically just one giant regular expression: https://github.com/falk-hueffner/metric-cooking/blob/master/...

The use case is a bit different (the task includes finding ingredients mentioned in a longer text, and it'll rather not parse something rather than parsing it wrong), but it works fairly OK and can even parse things like "3/4 cup plus 2 tablespoons packed light-brown sugar".

Comment · 5 years ago

For the quite similar puzzle Atomix, it also seems like the branching factor would be much higher for backward search because upper bounds are weaker, but you can show that on average the branching factor is actually the same [1]. I wonder if the same argument would work here.

[1] http://hueffner.de/falk/hueffner-studienarbeit-atomix.pdf Section 5.5

Comment · 5 years ago

You don't necessarily need a wider type (which might be slower to work with), you can just calculate (n|1) * ((n+1)/2). Clang does something much more complicated, probably because for it is just a special case of some much more general optimization.

Comment · 5 years ago

Slightly simpler:

    c = x ^ y;
    return popcount(((c >> 1) | 0x8080808080808080) - c) & 0x8080808080808080);